OpenRouter delegation — the any-model intern

ask-or is the same OpenRouter plumbing as ask-deepseek, but model-agnostic: you choose the slug. Treat it as an intern you can swap brains on — pick the cheapest model that clears the bar, or a frontier model when the task is hard. Same senior/intern discipline: scope it, hand it the material, review before shipping.

ask-deepseek is just the DeepSeek-default preset of this; ask-or exposes the whole OpenRouter catalogue.

When to reach for ask-or (vs ask-deepseek)

• You want a specific non-DeepSeek model (a different family for a second opinion, a frontier model for a tricky prompt, a cheap open model for bulk).
• You're comparing models on the same prompt.
• Otherwise, for generic cheap text work, /deepseek is the simpler default.

Tool

ask-or (at ~/.claude/bin/ask-or). Stdlib Python. Model is required — via -m SLUG or OPENROUTER_MODEL env; there is no silent default, so you never spend on a model you didn't choose. Needs OPENROUTER_API_KEY.

# pick any model explicitly
ask-or -m openai/gpt-5 "review this API design tradeoff: ..."
ask-or -m anthropic/claude-opus-4 -f spec.md "poke holes in this plan"
ask-or -m google/gemini-3-pro "summarize: ..."

# cheap open model for bulk
ask-or -m qwen/qwen3-max "rewrite formally: ..."

# set a default once, then omit -m
export OPENROUTER_MODEL=openai/gpt-5
ask-or "quick question"

# thinking mode (if the model supports it) + self-consistency vote
ask-or -m openai/gpt-5 -r high "hard logic problem"
ask-or -m google/gemini-3-pro -c 5 "Capital of Australia? city only"

# fan out one model over many prompts (shared cached prefix)
printf 'tldr A\ntldr B\n' | ask-or-batch -m qwen/qwen3-max -j 8

Flags: -m SLUG (or env OPENROUTER_MODEL; required), -s SYSTEM, -f FILE, -r [high|xhigh] (thinking mode, if supported), -c N (self-consistency vote), -t TEMP, --max-tokens N (env OPENROUTER_MAX_TOKENS), --json, --show-thinking, --timeout SEC (env OPENROUTER_TIMEOUT), -q. ask-or-batch adds -d DELIM, -j N; keep -s/-c identical for cache hits.

Agentic coding mode (--agentic)

--agentic drives your chosen model through the Codex harness (reads/edits a repo) — agentic coding on ANY OpenRouter model, no ChatGPT login. Internally shells ask-codex --or-model <your -m slug>.

ask-or -m anthropic/claude-opus-4 --agentic --apply -C ../wt "refactor X"
ask-or -m qwen/qwen3-max --agentic "review this repo's error handling"  # read-only

--apply edits files (use a throwaway worktree + review the diff); --cd DIR = root. Token-heavy (codex harness overhead). Tool-use reliability varies by model — stronger models drive the agent better.

The intern protocol (same as the others)

1. Brief — the intern has no repo memory; bundle context with -f. For "is X implemented?" feed code, not docs.
2. Assign — pick the model: cheap/open for bulk, frontier for hard prompts.
3. Review (mandatory) — spot-check; models hallucinate. Use -c N only for short verifiable answers.
4. Integrate — use the validated result; say which model produced it.

When NOT to use

• DeepSeek specifically is fine → /deepseek (simpler, cheap default).
• Image/video → /gemini. Repo coding/edits → /codex.
• Final correctness calls → Claude, never blind.

Notes

• Same OpenRouter account/key as deepseek. Caching is automatic (~0.25x cached).
• Pricing varies wildly by model — frontier slugs cost much more than DeepSeek. Match the model to the task.
• Related: [[reference-deepseek-delegation]], the gemini and codex skills — one KULI worker pool, same review discipline.